Frequency-decoupled dynamic graph learning with degradation prior for robust state-of-health estimation of lithium-ion batteries

Abstract

State-of-health (SOH) estimation of lithium-ion batteries is critical for reliable battery management in energy storage systems, where SOH must be inferred from noisy multi-frequency health indicators under varying operating conditions. Existing methods often fail to decouple multi-frequency degradation information, capture dynamic couplings among health indicators, and preserve consistency with degradation mechanisms, which limits their robustness across temperatures and chemistries. To address these issues, this study proposes a degradation-prior-enhanced frequency-decoupled dynamic graph learning framework for robust SOH estimation. STL decomposition is first employed to separate low-frequency degradation trends from high-frequency fluctuations, each handled by a dedicated temporal branch. A degradation-guided dynamic graph module is then developed, integrating health-indicator coupling priors and degradation-direction consistency priors to suppress spurious correlations. Finally, degradation-guided loss terms are embedded to encourage SOH trajectories consistent with battery aging behavior. Experiments on three public datasets (NASA, CALCE, and Zhu et al. dataset) show that the proposed method achieves an average mean absolute percentage error (MAPE) of 0.798%, outperforming mainstream baselines. In the cross-temperature transfer experiments, compared with the best-performing baseline GRU, the proposed method reduces the average root mean square error (RMSE) and mean absolute percentage error (MAPE) by 57.41% and 60.03%, respectively, indicating greater robustness and more reliable support for battery health management in practical battery monitoring.